Abraham · Independent AI research
Intelligence is
not learned alone.
Controlled experiments in how models learn from other models: inherited knowledge, verified teaching, and specialized fine-tuning.
This is the program’s motivating thesis—not a claim that the completed evidence has already established it.
Research engineering, in practice
From a hypothesis to an auditable experiment.
Paired experimental designs. Verified synthetic curricula. LoRA training and learning curves. Independent scoring and provenance checks. The reports document the implementation choices, measured outcomes, and failures that shaped the next study.
Explore the work behind the resultsLatest completed report · ADAPTIVE-APPRENTICESHIP
Adaptive Apprenticeship: Small Average Gain, Inconsistent Lineages
Static teaching scored 172/384 (44.79%); adaptive teaching 180/384 (46.88%). The paired mean difference was +2.08 percentage points, but only one of three lineages improved.
Research catalogue
Completed reports
06Adaptive Apprenticeship: Small Average Gain, Inconsistent LineagesCan student-feedback-guided teaching improve a specialized model more than matched-dose static teaching?Exploratory adaptive fine-tuning studyADAPTIVE-APPRENTICESHIP2026Report · PDF07Teaching and Specialization: The Pilot ArchiveWhat did earlier tests of rules, examples, explanation supervision and feedback reveal about model specialization?Ten exploratory pilots and diagnosticsTEACHING-PILOT-ARCHIVE2026Report · PDF05Route Apprenticeship: Small Gains, Incomplete AnswersDo model-written explanations improve a student more than reference explanations of the same verified answers?Exploratory fine-tuning studyROUTE-APPRENTICESHIP2026Report · PDF04More Fitting, Limited TransferDoes more answer-only fine-tuning improve fitting and transfer to new computations?Exploratory fitting studyFITTING-R22026Report · PDF03When the Student Does Not ActWas the student able to execute the action interface needed for a four-condition teaching comparison?Feasibility failure reportAPPRENTICESHIP-R32026Report · PDF02Cumulative Machine CultureCan knowledge accumulate across generations of a frozen language model when the only channel between generations is text?Completed research reportE0022026Report · PDF01Kleos Framing: A Confirmatory StudyDoes framing excellence as enduring renown change objective reasoning accuracy beyond closely matched social attribution?Completed research reportE001c2026Report · Note · PDFResearch program
01Model ApprenticeshipCan student feedback help a stronger teacher build a better specialized model?In preparation · Feedback follow-upAPPRENTICESHIP2026ProgramMethod before mythology
A laboratory that publishes its constraints.
Questions are falsifiable. Confirmatory decisions are frozen before execution. Negative results remain results. Claims stop where the sampled models, tasks, and procedures stop.
Read the methods & about